[BioC] Bug in maNormNN

[Tarca, Adi]

Hi Elizabeth,
Thanks for sending me the data, so I was able to see where the problem
was. The maNormNN function uses the nnet function to fit a spatial and
intensity based normalization model. Sometimes nnet function converges
without finding a proper solution. Although maNormNN was designed to
deal with such rare situations by restarting again the neural network
training, with your particular data set that was not sufficient to
ensure finding a good solution.
Now I have fixed this problem by increasing the number of attempts to
find a solution before giving up, and in the very improbable situation
that such convergence problem occurs an explicit warning message will be
given.
This fix should appear soon in the devel version of the nnNorm library
on the bioConductor website, as I have already committed the change in
the package. 

Thanks again,
Laurentiu

-----Original Message-----
From: Elizabeth Thomas [mailto:ethomas at CGR.Harvard.edu] 
Sent: Friday, November 11, 2005 2:05 PM
To: 'ltarca at rsvs.ulaval.ca'
Subject: Bug in maNormNN

Hello,

I need to a spatial normalization of some microarray data for a project
that
I'm working on, and after some research on different methods, I decided
to
use two methods which have been implemented in R and are available as
part
of the Bioconductor library, namely the maNorm scaled print-tip loess
method, and your neural network method.  I am seeing some odd artifacts
in
the output from the maNormNN, and I was wondering if you've seen this
before.  I wrote a little visualization program for mapping the M log
ratio
expression values onto the array layouts, and it demonstrates the
problem.
Here's an example where everything is working as it should:

Before normalization
http://eli-nati.fletzet.com/eli/ArrayLayouts/Normalization/Gasch01y12/ga
sch0
1y12_noNorm.8525.xls.png
Print tip intensity dependent (geonorm1)
http://eli-nati.fletzet.com/eli/ArrayLayouts/Normalization/Gasch01y12/ga
sch0
1y12_geoNorm1.8525.xls.png
Neural Network intensity dependent (geonorm2)
http://eli-nati.fletzet.com/eli/ArrayLayouts/Normalization/Gasch01y12/ga
sch0
1y12_geoNorm2.8525.xls.png

And here's an example where something has gone wrong with geonorm2
(maNormNN):
Before normalization
http://eli-nati.fletzet.com/eli/ArrayLayouts/Normalization/Gasch01y12/ga
sch0
1y12_noNorm.1010.xls.png
Print tip intensity dependent (geonorm1)
http://eli-nati.fletzet.com/eli/ArrayLayouts/Normalization/Gasch01y12/ga
sch0
1y12_geoNorm1.1010.xls.png
Neural Network intensity dependent (geonorm2)
http://eli-nati.fletzet.com/eli/ArrayLayouts/Normalization/Gasch01y12/ga
sch0
1y12_geoNorm2.1010.xls.png

The problem is reproduceable, and as you can see it doesn't affect the
geonorm1 results, which go through almost identical processing (maNorm=s
rather than maNormNN, but otherwise the same).  
Have you seen anything like this before?  Any tips?  I can send you the
original data and exactly what I did.  Generally, the maNormNN method
works
better than maNorm, particularly for data sets where the block size is
large
and most errors are within a block rather than across blocks.  As
advertised.  However, if these quirky artifacts keep popping up, then I
won't be able to trust maNormNN for my analyses.

thanks,
Elizabeth

------------------------------------------------------------------------
--------